Skip to main content
Advertisement
Browse Subject Areas
?

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

< Back to Article

Fig 1.

Example of a dynamic contrast-enhanced (DCE) MRI sequence for assessment of HCC enhancement.

The first row presents the pre-TACE examination, while the second row presents the post-TACE examination. The white arrow indicates the location of the HCC across the arterial phases before and after treatment.

More »

Fig 1 Expand

Fig 2.

Future post-TACE DCE-MRI prediction framework of HCC patients.

At training, feature maps are extracted from the baseline (pre-TACE) DCE-MRI sequences using a backbone ResNet, and projected in latent space, with data points partitioned within in a discriminant graph, with viable (V), equivocal (E) and nonviable (NV) tumor classes. Latent space samples are processed with three separate branches: a temporal branch modeling the dynamic changes, a structural branch capturing the morphological properties within single volumes, and global branch integrating an adversarial domain-translation loss, measuring the cost of mapping between source (pre-TACE ) and target (post-TACE ) domains in latent space , with the three branches combined to provide complementary features for inference. At inference time, the pre-treatment DCE-MRI is processed through a BatchNorm (BN) layer and a decoder followed by a softmax layer, to generate as an output, the follow-up sequence describing HCC response to TACE.

More »

Fig 2 Expand

Fig 3.

Schematic representation of the the temporal branch, where a series of patches are extracted from the input volumes and features partitioned to generate a total of T × P patches for a particular input sequence.

Graph nodes are constructed from temporal feature patches, providing a representation of the entire sequence such as , which are processed with a discriminant graph network, followed by a max pool operation to obtain the feature vector.

More »

Fig 3 Expand

Fig 4.

Schematic representation of the structural branch, where spatial relations between extracted patches from the single enhanced volumes are exploited in order to represent the morphological features from the DCE-MRI sequence.

Graph nodes stemming from single volume feature patches are combined together in the discriminant graph network to form the feature vector of the structural data.

More »

Fig 4 Expand

Fig 5.

Flowchart of patient selection.

More »

Fig 5 Expand

Fig 6.

Illustration of deformable registration results on 4 DCE-MRI cases (columns), achieving the non-rigid alignment of pre-TACE to post-TACE DCE-MRI examinations used for training purposes.

The last row illustrates the deformation vector fields (DVF) obtained on the coronal slices.

More »

Fig 6 Expand

Fig 7.

Confusion matrices of tumor viability classification (V: Viable, E: Equivocal, NV: Non-viable) of HCC following TACE based on the LI-RADS score.

Inputs to each model were the annotated pre-treatment DCE-MRI acquisitions. (a) ResCNN [19]; (b) RadiomicsDL [20]; (c) Proposed STDGNN method.

More »

Fig 7 Expand

Table 1.

Ablation experiments, comparing the addition of the structural and temporal branches to the proposed STDGNN.

A comparison is also made between the fully convolutional network (FCN) and the graph convolutional network (GCN). DSC: Dice similarity in %, MSE: Mean squared error (mm), Acc: classification accuracy in %, AUC: area under curve (AUC). Analysis of differences was performed by paired Wilcoxon tests (p < 0.05). Bold values indicate significant difference to baseline.

More »

Table 1 Expand

Fig 8.

Box-plots of the HCC segmentations from the predicted post-TACE images based on tumor viability class, using the original post-TACE examinations as the reference, and compared to 4 generation methods.

More »

Fig 8 Expand

Fig 9.

Pre-TACE (used as input) and post-TACE images with HCC segmentations generated through the decoder for 4 different liver cancer patients.

The first row depicts the pre-TACE examination with HCC delineation used as the baseline input. The second row presents the ground-truth post-TACE examinations, while the third row presents the predicted post-TACE images.

More »

Fig 9 Expand

Table 2.

Comparison with predictive methods.

DSC: Dice score coefficient of HCC segmentation, HD: Hausdorff distance (mm) of HCC segmentation, MSE: Mean squared error in voxel intensities of the predicted images. Analysis of differences was performed by paired Wilcoxon tests (p < 0.05). Bold values indicate significant difference.

More »

Table 2 Expand

Fig 10.

(a) Predicted tumor enhancement at sequential time points from the predicted post-TACE DCE-MRI images. Color maps indicate the changes within HCC tissue area. (b) Sample perfusion curves extracted from different liver regions (aorta, liver, HCC, portal vein), with the dashed lines indicating ground-truth post-TACE images.

More »

Fig 10 Expand

Table 3.

Parametric and non-parametric perfusion values extracted from the ground-truth (GT) and proposed predictive model (STDGNN).

Analysis of differences was performed by paired Wilcoxon tests (p < 0.05). Bold values indicate significant difference.

More »

Table 3 Expand